Add Apple Silicon support via vllm-metal

[ereid7]

Summary

• Adds Apple Silicon (macOS) support via vllm-metal — same plugin, same API, no code fork
• 3-line fix in processor.py: always call .cpu() before .numpy() so MPS tensors convert correctly
• README updated with full Apple Silicon setup guide, Open WebUI standalone instructions, and a note about the required vllm-metal PR #124

What changed

src/qr_sampler/processor.py — The _to_numpy() method previously only called .cpu() for CUDA tensors. MPS (Metal) tensors hit the else branch where .numpy() fails because the tensor isn't on CPU. Fix: always call .cpu() before .numpy() (no-op on CPU tensors).

README.md — Added:

• Apple Silicon setup section with MLX-format model example (mlx-community/Qwen3-0.6B-4bit)
• Verification step (check server logs for QRSamplerLogitsProcessor initialized)
• Both /v1/completions and /v1/chat/completions curl examples
• Prerequisite note about vllm-metal PR #124 (required for plugin discovery)
• Open WebUI standalone Docker instructions for Apple Silicon
• Split Web UI section into NVIDIA/Linux and Apple Silicon paths

.gitignore — Added .webui_secret_key (generated by Open WebUI, should not be committed)

Testing

• 308/308 unit tests pass
• E2E verified on Apple Silicon: both /v1/completions and /v1/chat/completions return entropy-driven responses
• Per-token sampling logs confirmed — full pipeline active (z-score, u-value, token selection via CDF)
• Open WebUI verified working against vllm-metal server

Dependencies

Requires vllm-metal PR #124 to be merged for plugin discovery to work. Without it, vllm-metal silently skips custom logits processors. The PR is a 9-line patch that mirrors GPUModelRunner's existing pattern.

[alchemystack]

Great, thanks!